Lilith.
⌕
Editorial illustration: Daybreak Red gets a containment architecture for supervised cyber tests
Lilith illustration · editorial remix

OpenAI has described how to deploy Daybreak Red for supervised security investigations. Codex runs as an unprivileged user on a separate agent VM, the network permits access only to the approved target, and an independent store records tool activity, requests to the test application and sign-in events.

Codex works inside a lab whose walls it cannot move

An analyst directs the architecture from a workstation. Commands execute on a separate VM in an isolated VPC, while the target should be a non-production application populated with synthetic data. OpenAI recommends a dedicated non-admin account, short-lived credentials and network rules enforced outside Codex itself.

Before an investigation, the team tests writes inside the approved workspace, blocked writes outside it, access to an allowed URL and rejection of a destination outside the allowlist. Logs should live in storage that the agent account cannot alter or delete.

Safety moves out of the prompt and into infrastructure

The most important part of the guide is not a model prompt. It is the separation of authority among the analyst, agent VM, firewall and independent telemetry. For a security team, approval of a sensitive action is only one layer. Network controls, operating-system permissions and an external way to stop the VM must enforce the investigation scope too.

The pattern extends beyond Daybreak Red. Once an agent can use a shell, browser, connectors or MCP tools, every route needs its own access control. A single allowlist in the model client offers little protection if another tool provides a path out.

The reference guide tests the setup, not the product's safety

This is a vendor-authored starting point, not an independent audit. Its walkthrough reproduces CVE-2026-117649 on a disposable Gitea 28.1.0 instance and limits the exercise to a small set of predefined GET requests. That can validate configuration and traceability for one run, but it says little about an unknown vulnerability, an active adversary or a misconfigured cloud account.

Auto-review also governs only the actions routed to it. Teams still need to verify the sandbox, firewall, browsers, connectors and log retention separately.

Complete traces and a fast kill switch will decide whether it works

Operational readiness requires a drill in which an operator stops the VM from outside, revokes temporary credentials and matches every command to a target request in independent logs. The useful signal will be the number of gaps between Codex events, network telemetry and application records.

Teams should also track incorrect approvals, out-of-scope access attempts and the time required to stop a run. Without those measurements, the architecture remains a careful floor plan rather than an operating security system.

Lilith's verdict

A capable cyber agent belongs in a lab where a human holds the kill switch and the camera sits behind glass. OpenAI is finally drawing the cage, not merely improving the creature inside it.